HolosAccent

Indlela iHolosAccent Eguqula Izwakala Yakho Ibe Isikali Sokuphimisa

Uma ufunda ibinzana ngokukhuluma, kukhona imibuzo eminingana ebalulekile: **yiziphi izwi ozikhiqizile, yiziphi amagama owaphelele, futhi umusho uhamba kanjani?** I-pipeline entsha yokulinganisa yeHolosAccent ihlanganisa ubufakazi ngalezi ezintathu ukuze ikhiqize isikali

HolosAccent· ·4 min read
✦✦✦ Murzik

Translated from English · Read the original

Ukulungiselela Ukuqopha kwakho

Okokuqala, umsindo uqoshwa ube ifomethi ejwayelekile: umsindo owodwa othathwe ku-16 kHz.

Isistimu ibala isilinganiso se-signal-to-noise sesiqophi. Uma lesi silinganiso singaphansi kuka 25 dB futhi ukuhlelwa komsindo kutholakala, isebenzisa i-filter yokunciphisa umsindo ngaphambi kokuhlaziya. Uma ukuhlelwa kwehluleka, iyaqhubeka nokuqopha kwangempela.

I-waveform ibuye ilungiswe ukuze ihambisane nezidingo zokufaka zemodeli yokuhlaziya inkulumo. Ukuthula kanye nokuqopha okungaphansi kwesigamu sesekhondi kuthola isikali esingu-zero.

Ukuvumela Ukuphimisa Okwehlukile

Okulandelayo, umbhalo wokubhekisela uguqulwa ube ukuphimisa kwefonetiki usebenzisa i-engine yokuphimisa. Igama lingaba nezinhlobo eziningi ezamukelekayo.

Isethi yokuphimisa iqukethe:

  • Izinhlobo zaseBrithani ne zaseMelika.
  • Izimo ezivamile ezinciphisiwe, ezifana ne-vowel engacindezelwanga ku-to.
  • Okukhethwe okunye okusekelwe kumongo.

Isibonelo, i-consonant yokugcina ku-close the door ihlukile kunaleyo ku-a close friend. Isistimu ingakhetha phakathi kwezinketho ezisekelwayo njengoba icubungula umusho.

Ukulalela Amazwi, hhayi Ukwakha Kabusha Amagama

Ukuqopha kudlula kumodeli ye-acoustic. Kukhishwa ubufakazi bamazwi ahlukene enkulumo ngokuhamba kwesikhathi, cishe ngokuhlukaniswa kuka-20-millisecond.

Lokhu kuyimephu ye-fonetiki yokuqopha. Isikali asisebenzisi imodeli yolimi ezingeni lamagama ukuze sakhe kabusha umusho.

Isistimu ibuye ihlole lezo ziphumo ze-acoustic ngezindlela ezimbili:

Indlela Okukwenza
Ukulinganisa komsindo okulindelekile Ithola isikhathi esifanele kakhulu samazwi alindelekile, ikhetha uhlobo oluphelele lokuphimisa lwamagama ngamunye.
Ukubonwa komsindo okuzimele Ilandela amazwi athandwa yimodeli ngaphandle kokufuna umusho wokubhekisela.

Zombili lezi zindlela ziyadingeka ngoba ukulinganisa kungabeka umsindo okulindelekile endaweni ethile yokuqopha noma umkhulumeli engawukaleli. Ukubonwa komsindo okuzimele kunikeza ubufakazi obengeziwe mayelana nalokho okukhulunywe ngempela.

Ukufanisa futhi kuthatha isikhathi ngokubheka, okwenza kube nzima ukuthi umsindo kwigama elilodwa ubale njengobufakazi begama elikude.

Ukulinganisa Umsindo Ngamunye

Kuwo wonke umsindo olinganiswe, isikali sibala isilinganiso sokunemba kokuphimisa: ukuthi imodeli isekela kanjani umsindo okulindelekile uma kuqhathaniswa namazwi ayinqaba ngalesi sikhathi.

Futhi ibheka:

  • I-probability ephezulu yomsindo okulindelekile.
  • Ukuba ukubonwa komsindo okuzimele kuthole umsindo ofane eduze.

Ukumiswa kwezibalo okwenziwe kushintsha lobu bufakazi babe inani lokunemba komsindo eliphakathi kuka-0 no-100.

Umsindo ongatholakali kumphumela wokubonwa okuzimele awutholi ngokuzenzakalelayo i-zero. Ubufakazi be-acoustic busasebenza.

Ukunquma Amagama Adinga Ukuqaphela

Impendulo yamagama ihlanganisa la manani omsindo nezihlolo ezicacile zamaphutha. Igama kufanele:

  • Ibe nobufakazi obanele bokuthi ikhulunywe.
  • Ihlangabezane nesilinganiso esiphakathi esimiswe.
  • Igcine ekushintsheni, ukufakwa, noma amaphutha amakhulu okushoda komsindo atholakele.

Ukushintsha okucacile kudinga ukusekelwa okuncane kumsindo okulindelekile kanye nokusekelwa okuqinile kumsindo okuphikisanayo.

Ubufakazi Isilinganiso samanje
Ukusekelwa kumsindo okulindelekile Ngaphansi kuka 0.05
Ukusekelwa kokufaka kwaso Okungenani 0.5

Lezi zilinganiso zobufakazi bemodeli, hhayi ama-probability ukuthi umlaleli womuntu azobona igama lingalungile.

Amagama aboniswe, isistimu ibuyisela ukuphimisa okulindelekile kanye nokulinganisa kwayo kwamazwi alalwe, okubhalwe ku-International Phonetic Alphabet (IPA). Lokho kungaba kungaphelele, ikakhulukazi eduze kwemiphetho yamagama.

Ukwakha Isikali Somusho

Isikali somusho siqala ngokuhlanganiswa okwenziwe kwenani elilandelayo:

  • Inani lokunemba komsindo eliphakathi.
  • Isabelo samazwi alindelekile afaniswe ukubonwa komsindo okuzimele.
  • Inani lamazwi enkulumo aphimiswa ngomzuzwana.
  • Ukuhoxiswa okude phakathi kwamagama.

Ukubalwa kwesikhathi akubali ukuthula ngaphambi komsindo wokuqala olinganiswe nangemva kokugcina.

Lezi zici zifaka isandla ku-isikali esisodwa esiphelele. Ukuqaliswa kwamanje akuhlinzeki ngezilinganiso ezihlukene ezibhekiswe kumsebenzisi ezifana ne-fluency, completeness, noma intonation.

Indlela i-Word Credit Ethinta Ngayo Umphumela

Isikali bese sisebenzisa i-multiplier ye-word-credit ecacile. Igama ngalinye lilayishwa ngenani lamazwi ekuphimiseni okukhethiwe.

Umphumela wegama I-Credit etholwe
Kulungile Isisindo esigcwele
Kukhulunywe kodwa kubonisiwe Isigamu sesine sesisindo sayo
Kushiywe Akukho credit

Ukubalwa kwalo kungukuthi:

word credit =
    (correct sound weight + 0.25 × flagged spoken sound weight)
    / total expected sound weight

raw score =
    sentence estimate, clipped to 0–100, × word credit

Ngenxa yalokho, igama elide elingalungile ngokuvamile libiza kakhulu kunelincane. Ukuze ufunde ingxenye encane yomusho nakho kunciphisa isikali, noma leyo ngxenye icacile.

Ukuqhathanisa Isikali Sakho Nombhalo Wokubhekisela

Ekugcineni, isistimu ingakwazi ukujwayela isikali esiyinqaba isebenzisa ukuqopha okubhekiselwe kwengxenye efanayo.

Ilinganisa i-reference eyakhiwe nge-version efanayo yesikali, futhi uma i-baseline efanelekile isivele ikhona, isebenzisa:

displayed score = min(100, raw score / reference score × 100)

Isibonelo, isikali esiyinqaba esingu-80 uma siqhathaniswa nesikali sokubhekisela esingu-88 siba cishe 90.9.

Lokhu kulungisa kubheka iqiniso lokuthi imodeli inganikeza ngisho nokuqopha okubhekiselwe ngaphansi kuka-100.

Ama-baselines axhumene ne:

  • Izwi lokubhekisela nemodeli yokukhiqiza inkulumo.
  • Imodeli yokulinganisa.
  • I-vocabulary.
  • Ukuhlolwa kanye nohlobo lwemithetho.

Uma i-baseline efanelekile ingatholakali noma ingasebenziseki, umphumela ugcina isikali saso esiyinqaba. Umzamo wokuqala ungahlala ungashintshiwe ngenkathi i-baseline ilungiselelwa imizamo ezayo.

Okusho Isikali Sakho

Inombolo ephumayo iy isikali sokuzilolonga esikalini seHolosAccent.

Isikali esingu-90 asisho ukuthi amaphesenti angu-90 wamazwi akho alungile, futhi asisho izinga le-CEFR noma ukufana nenkulumo yomuntu owaziyo ulimi.

Ukumiswa kwenziwe kusetshenziswa ukuqopha okwenziwe ngochwepheshe, kodwa uhlelo olusha lokulinganisa luseyizivivinyo futhi izimiso zalo zokwamukelwa ngokuphelele azikakafakazelwa. Lungaphuthelwa amaphutha noma libonise inkulumo evumelekile.

Impendulo yamagama ngamunye kanye namaphethini phakathi kwemizamo ephindaphindiwe kunikeza umongo wokuchaza le nombolo.

HolosAccent

Hear the difference. Fix it in two minutes a day.

Read a phrase out loud and HolosAccent scores every sound — then shows exactly which one to fix. Murzik grows with you.

Help improve the HolosAccent blog?

With your permission, HolosAccent measures visits, articles and languages read, active reading time, scrolling, and app-link clicks. We use a random visitor ID and record your IP address, approximate country, browser/device category, referrer hostname, and campaign labels. IP addresses are removed after 30 days; other analytics and consent records expire after 180 days. We do not use advertising trackers or connect these records to your app account.

Analytics is optional. You can refuse, withdraw consent, export, or delete this browser's analytics at any time here. Access to the blog stays the same. Contact: support@holosaccent.app.

Privacy · GDPR